Inexact Multi-Task Learning for Fetal Anastomoses Detection

Alessandro Casella, Gaia Romana De Paolis, Elena De Momi, Dario Paladini, Sara Moccia, Leonardo S. Mattos

Abstract:

Twin-to-Twin Transfusion Syndrome (TTTS) is a rare
pathology that may affect monochorionic twin pregnancies. TTTS depends on the unbalanced blood transfer from
one twin (the donor) to the other (the recipient) through
abnormal placental vascular anastomoses. Currently, the
treatment for TTTS consists of the photo-ablation of
abnormal anastomoses in fetoscopic laser surgery [1].
Residual anastomoses still represent a major complication [2] and their identification is not a trivial task.
Visual challenges such as small field of view, amniotic
fluid turbidity, low-resolution imaging, and unfavourable
views are due to the position of the insertion site for
the tools. To support surgeons, researchers are working
on vessel and placenta segmentation [3], [4]. Recently,
[5] presented the first multi-centre large-scale dataset to
improve the current state-of-the-art in segmentation and
registration in fetoscopy. However, to date, there is no
work in the literature on anastomosis detection. There is
also no available datasets for this task.
This work aims to develop a deep-learning-based framework for anastomosis detection in intra-operative fetoscopic videos from inexact labels. Considering the challenges of labelling anastomoses, we propose a weaklysupervised strategy by training a multi-task convolutional
neural network (CNN) for (i) segmenting vessels in the
fetoscopy frame and (ii) classifying frames as containing
anastomoses or not. Relying on class activation mapping
(CAM), anastomosis detection is then accomplished.